This is the code repository for paper: Ubi-SleepNet: Advanced Multimodal Fusion Techniques for Three-stage Sleep Classification Using Ubiquitous Sensing.
- Data Pre-processing code is available at:MakeSenseOfSleep which should produce a h5 file for MESA dataset.
- This repository includes the builders for MESA statistic features and Apple Watch dataset with statistic features
- For the MESA with HRV feature set, please go to MakeSenseOfSleep. It includes a dataset builder that can build the HRV feature dataset.
- This repository also includes a data builder to build the dataset that uses the raw accelerometer and HR data collected from the Apple Watch dataset.
To ensure the experiments run smoothly, please create a python 3.8 environment, and please be aware, the pytables and h5py requires to be installed via conda .
you could run a non-attention based model by:
python -m train_val_test --nn_type Vggacc79f174_7 --epochs 20 --dataset mesa
To run the attention model, the modality should be specified. The code below is an example:
python -m train_val_test --nn_type VggAcc79F174_SplitModal_SANTimeDimMatrixAttOnMod1NLayer1 --epochs 20 --dataset mesa --att_on_modality act